Leak Sanitizer port to Windows

[clemenswasser]

Summary

Initial Leak Sanitizer port to Windows.
The main part of the port was to implement StopTheWorld for Windows.
Many tests of the lsan test suite seem to run fine (leaks are correctly being reported), however __lsan_ignore_object doesn't seem to work.
Some other tests that fail are: new_array_with_dtor_0 and large_allocation_leak.
The test suite for lsan won't run for me. Nearly every test fails with this error message:

$ "LSAN_BASE=use_stacks=0:use_registers=0"
'LSAN_BASE=use_stacks=0:use_registers=0': command not found

I also need to figure out, how to get clang to auto link the lsan runtime when specifying -fsanitize=leak or -fsanitize=address on Windows.

[clemenswasser]

Changed to core.autocrlf=false

[clemenswasser]

Move __lsan_init call inside of SANITIZER_WINDOWS ifdef

[clemenswasser]

Apply clang-format

[clemenswasser]

Use LF line endings

[MyDeveloperDay]

compiler-rt/lib/lsan/lsan.h

20

clang-format?

compiler-rt/lib/lsan/lsan_common.h

49

you probably don't want to change this do you?

[MyDeveloperDay]

You need to add some reviewers to this revision. check out the CODE_OWNERS.txt or what I sometimes do is to git log one of the files you are changing and see who is a major contributor.

[clemenswasser]

Remove wrong interceptors

[vitalybuka]

LGTM in general, but it can be cut into smaller patches.

+ @rnk for Windows suff

compiler-rt/lib/lsan/lsan.h

20

@clemenswasser Can you please a separate tiny patch which clang-format nearby lines

compiler-rt/lib/lsan/lsan_allocator.cpp

30 ↗

(On Diff #392017)

This also can go into a separate patch
BTW I would slightly prefer just UL -> ULL

compiler-rt/lib/lsan/lsan_common.cpp

502

I guess do needs this as well on windows
The problem is that we have no MemoryMappingLayout implementation for Windows.
I guess it can be done with VirtualQueryEx
for now can you just ifdef only body of the ProcessRootRegion with TODO?

compiler-rt/lib/lsan/lsan_common.h

49

I am fine like this, fighting clang-format is not useful
however it would be nice if you clang-format relevant block in a separate patch then then rebase this patch on top of it

compiler-rt/lib/lsan/lsan_common_win.cpp

49

how this can work without ProcessGlobalRegions?

compiler-rt/lib/lsan/lsan_interceptors.cpp

532

Can you please just convert unneeded stuff from INTERCEPT_FUNCTION into LSAN_MAYBE_INTERCEPT
In a separate patch

compiler-rt/lib/sanitizer_common/sanitizer_stoptheworld_win.cpp

4 ↗

(On Diff #392017)

sanitizer_stoptheworld_test.cpp needs to be ported to windows, probably with std::
can you please move this work and sanitizer_stoptheworld_win.cpp into separate patch/patches

[MyDeveloperDay]

compiler-rt/lib/lsan/lsan.h

20

Sorry, is there a reason we can't follow the current style in this patch? sorry did I miss something? Its not about fighting clang-format its about ensuring we follow the style from the lines around it which are following the clang-format style. I don't need you to clang-format the whole file (although that would be great because this directory has very low clang-format status)

https://clang.llvm.org/docs/ClangFormattedStatus.html

All I'm asking is that we don't indent the preprocessor directives to a non LLVM style.

i.e.

#include "lsan_win.h"
#if !SANITIZER_WINDOWS
#endif

vs

#  include "lsan_win.h"
#  if !SANITIZER_WINDOWS
#  endif

[MyDeveloperDay]

compiler-rt/lib/lsan/lsan_common.h

49

it might not seem useful to individuals, but I think its the guidance, and a consistent style really help new people to understand what's going on.

https://llvm.org/docs/Contributing.html#how-to-submit-a-patch

I'm a little surprise the "llvm-premerge-checks" didn't highlight this @kuhnel

[MyDeveloperDay]

oh! gosh I'm sorry I didn't realize that lsan has this non standard LLVM style .clang-format, my apologies! D100238: [sanitizer] Set IndentPPDirectives: AfterHash in .clang-format

kind of surprised TBH, doesn't feel the indentation happens enough to warrant it, but that's up to you its not my backyard.

$ grep "#if" *.cpp | wc -l
461
llvm-project/compiler-rt/lib/sanitizer_common
$ grep "# if" *.cpp | wc -l
27

[clemenswasser]

compiler-rt/lib/lsan/lsan.h

20

Yes, this is caused by clang-format. What should I do about it?

20

@vitalybuka Should I create a patch, where I format the whole file or just these #includes?

[MyDeveloperDay]

compiler-rt/lib/lsan/lsan.h

20

You can ignore my comment, whilst I think the lsan .clang-format file is incorrect in terms of matching the LLVM style it is correct from the lsan subprojects point of view. I see now that you are doing what is in the local .clang-format file and its the rest of the file is incorrect. (you would handle that separately or not at all)

The fact that the .clang-format file doesn't follow the LLVM style is unfortunate, but as I expressed thats upto you all.

[clemenswasser]

Split this patch up into multiple smaller ones:

• D115186
• D115204
• D115262

[vitalybuka]

In D115103#3176852, @clemenswasser wrote:

Split this patch up into multiple smaller ones:

• D115186
• D115204
• D115262

Please check "Edit Related Revisions" on the right of Phabricator. You can nicely chain this patches into "Stack"

In D115103#3176852, @clemenswasser wrote:

Split this patch up into multiple smaller ones:

• D115186
• D115204
• D115262

[clemenswasser]

Implement ProcessGlobalRegions
Fix Interceptors for Windows

[vitalybuka]

compiler-rt/lib/lsan/lsan_allocator.h

52–54

68–69

compiler-rt/lib/lsan/lsan_common.cpp

248–253

same

[clemenswasser]

Simplify Architecture checks

[clemenswasser]

@vitalybuka could you or someone else help me get the lsan tests running on Windows?
For example lets look at compiler-rt\test\lsan\TestCases\leak_check_at_exit.cpp (and basically all other tests, this serves as an easy example)
I imagine that the following defines a variable called LSAN_BASE :
// RUN: LSAN_BASE="use_stacks=0:use_registers=0"
This doesn't seem to work on Windows.
It gives the following error message:

[build] $ ":" "RUN: at line 2"
[build] $ "LSAN_BASE=use_stacks=0:use_registers=0"
[build] # command stderr:
[build] 'LSAN_BASE=use_stacks=0:use_registers=0': command not found

[vitalybuka]

In D115103#3207150, @clemenswasser wrote:

@vitalybuka could you or someone else help me get the lsan tests running on Windows?
For example lets look at compiler-rt\test\lsan\TestCases\leak_check_at_exit.cpp (and basically all other tests, this serves as an easy example)
I imagine that the following defines a variable called LSAN_BASE :
// RUN: LSAN_BASE="use_stacks=0:use_registers=0"
This doesn't seem to work on Windows.
It gives the following error message:

[build] $ ":" "RUN: at line 2"
[build] $ "LSAN_BASE=use_stacks=0:use_registers=0"
[build] # command stderr:
[build] 'LSAN_BASE=use_stacks=0:use_registers=0': command not found

Can you try to remove LSAN_BASE and just hardcode use_stacks=0:use_registers=0 ?
If it works don't mind if we do so for the rest (in a separate patch)

[clemenswasser]

The problem is, that LSAN_BASE is not always use_stacks=0:use_registers=0.
Also there are many test cases, where LSAN_BASE gets used multiple times.
Removing it would result in bad repetition.

[vitalybuka]

In D115103#3207210, @clemenswasser wrote:

The problem is, that LSAN_BASE is not always use_stacks=0:use_registers=0.

Off-cause , I am asking to replace with correct value each test.

Also there are many test cases, where LSAN_BASE gets used multiple times.
Removing it would result in bad repetition.

I don't like LSAN_BASE even without Windows.
Usually when test fail, I have a single command line to reproduce or put into debugger for just the step. Here I need to cherry-pick two line: LSAN_BASE= and actual step
also in most case it used just twice, so it's not a big loss in repetition

[clemenswasser]

@vitalybuka I have now removed all LSAN_BASE occurrences.
However the invocations of the compiled test executables don't seem to work either.
This doesn't work:
// RUN: %env_lsan_opts=use_stacks=0:use_registers=0 not %run %t foo 2>&1 | FileCheck %s --check-prefix=CHECK-do
It generates this error:

[build] $ ":" "RUN: at line 3"
[build] $ "env" "LSAN_OPTIONS=:detect_leaks=1:use_stacks=0:use_registers=0" "not" "E:\git\llvm-project\llvm\build\ninja_debug\projects\compiler-rt\test\lsan\X86_64LsanConfig\TestCases\Output\leak_check_at_exit.cpp.tmp.exe" "foo"
[build] note: command had no output on stdout or stderr
[build] error: command failed with exit status: True

One of the problems seems to be that %t doesn't have a .exe ending on Windows.
I also tried changing %t to %t.exe, but that doesn't fix the problem and produces the same error.
I think the not command(?) is the part that fails?

When I run it on my own, it does work:

PS E:\git\llvm-project> & 'E:\git\llvm-project\llvm\build\ninja_debug\projects\compiler-rt\test\lsan\X86_64LsanConfig\TestCases\Output\leak_check_at_exit.cpp.tmp.exe'
Test alloc: 000061A000000000.

=================================================================
==25768==ERROR: LeakSanitizer: detected memory leaks

Direct leak of 1337 byte(s) in 1 object(s) allocated from:
    #0 0x7ff6e2f745be in __asan_wrap_malloc E:\git\llvm-project\compiler-rt\lib\lsan\lsan_interceptors.cpp:75
    #1 0x7ff6e2f2aade in main E:\git\llvm-project\compiler-rt\test\lsan\TestCases\leak_check_at_exit.cpp:13:40
    #2 0x7ff6e2f79b5b in __scrt_common_main_seh d:\a01\_work\6\s\src\vctools\crt\vcstartup\src\startup\exe_common.inl:288
    #3 0x7fffde187033  (C:\WINDOWS\System32\KERNEL32.DLL+0x180017033)
    #4 0x7fffde8a2650  (C:\WINDOWS\SYSTEM32\ntdll.dll+0x180052650)

SUMMARY: LeakSanitizer: 1337 byte(s) leaked in 1 allocation(s).

[vitalybuka]

In D115103#3207867, @clemenswasser wrote:

@vitalybuka I have now removed all LSAN_BASE occurrences.
However the invocations of the compiled test executables don't seem to work either.
This doesn't work:
// RUN: %env_lsan_opts=use_stacks=0:use_registers=0 not %run %t foo 2>&1 | FileCheck %s --check-prefix=CHECK-do
It generates this error:

[build] $ ":" "RUN: at line 3"
[build] $ "env" "LSAN_OPTIONS=:detect_leaks=1:use_stacks=0:use_registers=0" "not" "E:\git\llvm-project\llvm\build\ninja_debug\projects\compiler-rt\test\lsan\X86_64LsanConfig\TestCases\Output\leak_check_at_exit.cpp.tmp.exe" "foo"
[build] note: command had no output on stdout or stderr
[build] error: command failed with exit status: True

does check-asan work for your setup?

[clemenswasser]

@vitalybuka
No check-asan doesn't work for me. It just hangs forever and does absolutely nothing. No output, nothing showing up in Task Manager with high CPU usage or anything.
However check-clang does work. Is there some documentation on running check-asan on windows.
I also saw that none of the Windows buildbots check/test anything related to sanitizers. It seems to me that none of the sanitizer stuff gets ever tested on Windows?
Do the test for sanitizers_common/asan/ubsan/etc. even work on Windows?

[vitalybuka]

In D115103#3209326, @clemenswasser wrote:

@vitalybuka
No check-asan doesn't work for me. It just hangs forever and does absolutely nothing. No output, nothing showing up in Task Manager with high CPU usage or anything.
However check-clang does work. Is there some documentation on running check-asan on windows.
I also saw that none of the Windows buildbots check/test anything related to sanitizers. It seems to me that none of the sanitizer stuff gets ever tested on Windows?
Do the test for sanitizers_common/asan/ubsan/etc. even work on Windows?

This one runs asan on Windows, you can try to replicated using logs. https://lab.llvm.org/buildbot/#/builders/127
I rarely work on Windows, but i did similar setup some time ago.

[rnk]

In D115103#3209468, @vitalybuka wrote:

In D115103#3209326, @clemenswasser wrote:

@vitalybuka
No check-asan doesn't work for me. It just hangs forever and does absolutely nothing. No output, nothing showing up in Task Manager with high CPU usage or anything.
However check-clang does work. Is there some documentation on running check-asan on windows.
I also saw that none of the Windows buildbots check/test anything related to sanitizers. It seems to me that none of the sanitizer stuff gets ever tested on Windows?
Do the test for sanitizers_common/asan/ubsan/etc. even work on Windows?

This one runs asan on Windows, you can try to replicated using logs. https://lab.llvm.org/buildbot/#/builders/127
I rarely work on Windows, but i did similar setup some time ago.

I will just confirm that check-asan is supposed to work, it worked for me a few months ago, and the linked buildbot tests it continuously.

However, WinASan is very sensitive to the OS and SDK version. The failure mode you described sounds plausible, something like a crash on startup that results in a message box that you can't see or something like that.

[rnk]

See a recent change rG7cd109b92c72855937273a6c8ab19016fbe27d33 by @aganea, who presumably was able to run the ASan tests for it.

[aganea]

It does work indeed. I am running under a "x64 Native Tools Command Prompt for VS 2019" prompt with MSVC 16.11.8.

D:\git\llvm-project>mkdir release && cd release
D:\git\llvm-project\release>cmake ../llvm -G Ninja -DLLVM_ENABLE_PROJECTS="clang;lld;llvm;compiler-rt" -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_PDB=ON -DLLVM_ENABLE_ASSERTIONS=ON
...
D:\git\llvm-project\release>ninja check-asan
[0/1] Running the AddressSanitizer tests
-- Testing: 669 tests, 16 workers --
Testing:  0.. 10.. 20.. 30.. 40.. 50.. 60.. 70.. 80.. 90..

Testing Time: 86.33s
  Unsupported      : 229
  Passed           : 426
  Expectedly Failed:  14

However you might need some Unix tools that don't come out-of-the-box on Windows.
You could first install the scoop package manager and then install any tool that you need:

D:\git\llvm-project>powershell
...
D:\git\llvm-project [main ↓9 +10 ~0 -0 !]> iwr -useb get.scoop.sh | iex
...
D:\git\llvm-project [main ↓9 +10 ~0 -0 !]> scoop install coreutils grep make sed unxutils sysinternals

And then run ninja check-asan again.

If you install sysinternals with scoop above, you can run procexp64 prior to ninja check-asan. That will give a much finer view on the running processes, as compared to the Task Manager. Let me know if I can help further.

[clemenswasser]

@aganea Thanks for your help. I already did everything you wrote (I am however using the coreutils provided by Git).
check-asan works on my new computer. I still don't know why it didn't work on my older one...
I will try to get the lsan tests working (and hopefully passing 😉) on windows soon.

[clemenswasser]

All the lsan tests, when run in lsan+asan mode, Deadlock in StopTheWord when creating the tracer thread...

-> check-lsan gets stuck, due to all the deadlocking tests :(

Could it be caused by asan intercepting the Win32 calls in `StopTheWorld?
@vitalybuka Could this be a problem? If so, how can I call the "real" functions, without being intercepted by asan?
Is there a way (for now), to only run the lsan tests in lsan only mode (no asan)?

[clemenswasser]

Updated the Tests.
The Problem with Lsan+Asan still exists. I have just disabled them for now.
I hope anyone can help me with getting the tests to run correctly?